Research Of Target Tracking Based On Sparse Representation

The sparse representation based object tracking method is one of the research hot spots in the field of computer vision,which is widely studied by researchers.However,it remains difficult to establish a strong robustness,high accuracy and real-time algorithm to solve all these cases where there exit illumination change,background clutter,partial occlusion,pose and scale change of the target in the sense of video sequences.So this paper probes into the sparse representation based object tracking methods in order to improving the robustness,accuracy and real-time of the algorithm.The main work of this paper is as follows:This paper presents a novel sparse representation model for object tracking based on the particle filter frame work framework.Due to the high computational complexity of the l₁ standard least squares sparse equation,which results in the low fps of the tracking process,all the algorithms in this paper are based on the multi task sparse representation l_2,1.In order to cope with the motion blur of the target,each frame is adjusted to a fixed size in advance.To enhance the robustness of the object tracking method,an object tracking algorithm based on multi task structure sparse appearance model is presented.Firstly,the algorithm makes the most use of the whole,local and spatial structural information by the means of the overlapping patches over samples,which is designed to ensure the accuracy of the tracking method.Then,the l_2,1 normalized least squares method is used to solve the sparse coding coefficients corresponding to each image block,and the similarity between the candidate samples and the target is determined by exploiting the alignment pooling algorithm,which can extract the coefficient information of the multi-task structure sparse representation.Finally,to reduce the probability of drift which comes from changes of the target appearance during the tracking,an adaptive template updating strategy based on incremental subspace learning and sparse representation is proposed.Since the single feature is not enough to deal with the variety of complex factors,a hybrid model visual tracking algorithm based on multi-feature fusion is proposed.It not only captures the latent spatial layout structure among the local patches from each target candidate via the local appearance model of intensity but also exploits the holistic appearance information among particles with the global templates of color histograms.To overcome the difficulties of arbitrary appearance changes,a dynamical updating strategy of target templates is presented.For the sake of resolving the frequently emerging outlier tasks in the resampling process and reduce the computational complexity of the target tracking algorithm,a hybrid model target tracking algorithm based on outlier rejection is explored.The method decomposes the representation matrix into two collaborative components,and the l_2,1 mixed-norm regularization term is imposed.